brintos

brintos / llvm-project-archived public Read only

0
0
Text · 19.8 KiB · db5d0fd Raw
523 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck --check-prefix=GFX9 %s3; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck --check-prefix=GFX7 %s4; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck --check-prefix=GFX10 %s5; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck --check-prefix=GFX11 %s6 7; FIXME:8; XUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck --check-prefix=GFX6 %s9 10define amdgpu_kernel void @store_lds_v3i32(ptr addrspace(3) %out, <3 x i32> %x) {11; GFX9-LABEL: store_lds_v3i32:12; GFX9:       ; %bb.0:13; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x1014; GFX9-NEXT:    s_waitcnt lgkmcnt(0)15; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x016; GFX9-NEXT:    ; kill: killed $sgpr4_sgpr517; GFX9-NEXT:    v_mov_b32_e32 v0, s018; GFX9-NEXT:    v_mov_b32_e32 v1, s119; GFX9-NEXT:    v_mov_b32_e32 v2, s220; GFX9-NEXT:    s_waitcnt lgkmcnt(0)21; GFX9-NEXT:    v_mov_b32_e32 v3, s322; GFX9-NEXT:    ds_write_b96 v3, v[0:2]23; GFX9-NEXT:    s_endpgm24;25; GFX7-LABEL: store_lds_v3i32:26; GFX7:       ; %bb.0:27; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x428; GFX7-NEXT:    s_waitcnt lgkmcnt(0)29; GFX7-NEXT:    s_load_dword s3, s[4:5], 0x030; GFX7-NEXT:    s_mov_b32 m0, -131; GFX7-NEXT:    v_mov_b32_e32 v0, s032; GFX7-NEXT:    v_mov_b32_e32 v1, s133; GFX7-NEXT:    v_mov_b32_e32 v2, s234; GFX7-NEXT:    s_waitcnt lgkmcnt(0)35; GFX7-NEXT:    v_mov_b32_e32 v3, s336; GFX7-NEXT:    ds_write_b96 v3, v[0:2]37; GFX7-NEXT:    s_endpgm38;39; GFX10-LABEL: store_lds_v3i32:40; GFX10:       ; %bb.0:41; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x1042; GFX10-NEXT:    s_waitcnt lgkmcnt(0)43; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x044; GFX10-NEXT:    ; kill: killed $sgpr4_sgpr545; GFX10-NEXT:    v_mov_b32_e32 v0, s046; GFX10-NEXT:    v_mov_b32_e32 v1, s147; GFX10-NEXT:    v_mov_b32_e32 v2, s248; GFX10-NEXT:    s_waitcnt lgkmcnt(0)49; GFX10-NEXT:    v_mov_b32_e32 v3, s350; GFX10-NEXT:    ds_write_b96 v3, v[0:2]51; GFX10-NEXT:    s_endpgm52;53; GFX11-LABEL: store_lds_v3i32:54; GFX11:       ; %bb.0:55; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x1056; GFX11-NEXT:    s_waitcnt lgkmcnt(0)57; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x058; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s159; GFX11-NEXT:    s_waitcnt lgkmcnt(0)60; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s361; GFX11-NEXT:    ds_store_b96 v3, v[0:2]62; GFX11-NEXT:    s_endpgm63  store <3 x i32> %x, ptr addrspace(3) %out64  ret void65}66 67define amdgpu_kernel void @store_lds_v3i32_align1(ptr addrspace(3) %out, <3 x i32> %x) {68; GFX9-LABEL: store_lds_v3i32_align1:69; GFX9:       ; %bb.0:70; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x1071; GFX9-NEXT:    s_waitcnt lgkmcnt(0)72; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x073; GFX9-NEXT:    ; kill: killed $sgpr4_sgpr574; GFX9-NEXT:    s_and_b32 s5, 0xffff, s075; GFX9-NEXT:    v_mov_b32_e32 v0, s076; GFX9-NEXT:    s_waitcnt lgkmcnt(0)77; GFX9-NEXT:    v_mov_b32_e32 v1, s378; GFX9-NEXT:    s_lshr_b32 s3, s5, 879; GFX9-NEXT:    s_lshr_b32 s4, s0, 1680; GFX9-NEXT:    ds_write_b8 v1, v081; GFX9-NEXT:    v_mov_b32_e32 v0, s382; GFX9-NEXT:    ds_write_b8 v1, v0 offset:183; GFX9-NEXT:    s_lshr_b32 s0, s0, 2484; GFX9-NEXT:    v_mov_b32_e32 v0, s485; GFX9-NEXT:    ds_write_b8 v1, v0 offset:286; GFX9-NEXT:    v_mov_b32_e32 v0, s087; GFX9-NEXT:    s_and_b32 s3, 0xffff, s188; GFX9-NEXT:    ds_write_b8 v1, v0 offset:389; GFX9-NEXT:    s_lshr_b32 s3, s3, 890; GFX9-NEXT:    v_mov_b32_e32 v0, s191; GFX9-NEXT:    s_lshr_b32 s0, s1, 1692; GFX9-NEXT:    ds_write_b8 v1, v0 offset:493; GFX9-NEXT:    v_mov_b32_e32 v0, s394; GFX9-NEXT:    ds_write_b8 v1, v0 offset:595; GFX9-NEXT:    s_lshr_b32 s1, s1, 2496; GFX9-NEXT:    v_mov_b32_e32 v0, s097; GFX9-NEXT:    ds_write_b8 v1, v0 offset:698; GFX9-NEXT:    v_mov_b32_e32 v0, s199; GFX9-NEXT:    s_and_b32 s1, 0xffff, s2100; GFX9-NEXT:    ds_write_b8 v1, v0 offset:7101; GFX9-NEXT:    s_lshr_b32 s1, s1, 8102; GFX9-NEXT:    v_mov_b32_e32 v0, s2103; GFX9-NEXT:    s_lshr_b32 s0, s2, 16104; GFX9-NEXT:    ds_write_b8 v1, v0 offset:8105; GFX9-NEXT:    v_mov_b32_e32 v0, s1106; GFX9-NEXT:    ds_write_b8 v1, v0 offset:9107; GFX9-NEXT:    s_lshr_b32 s1, s2, 24108; GFX9-NEXT:    v_mov_b32_e32 v0, s0109; GFX9-NEXT:    ds_write_b8 v1, v0 offset:10110; GFX9-NEXT:    v_mov_b32_e32 v0, s1111; GFX9-NEXT:    ds_write_b8 v1, v0 offset:11112; GFX9-NEXT:    s_endpgm113;114; GFX7-LABEL: store_lds_v3i32_align1:115; GFX7:       ; %bb.0:116; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x4117; GFX7-NEXT:    s_waitcnt lgkmcnt(0)118; GFX7-NEXT:    s_load_dword s3, s[4:5], 0x0119; GFX7-NEXT:    s_mov_b32 m0, -1120; GFX7-NEXT:    s_bfe_u32 s5, s0, 0x80008121; GFX7-NEXT:    v_mov_b32_e32 v0, s0122; GFX7-NEXT:    s_waitcnt lgkmcnt(0)123; GFX7-NEXT:    v_mov_b32_e32 v1, s3124; GFX7-NEXT:    s_lshr_b32 s4, s0, 16125; GFX7-NEXT:    ds_write_b8 v1, v0126; GFX7-NEXT:    v_mov_b32_e32 v0, s5127; GFX7-NEXT:    ds_write_b8 v1, v0 offset:1128; GFX7-NEXT:    s_lshr_b32 s0, s0, 24129; GFX7-NEXT:    v_mov_b32_e32 v0, s4130; GFX7-NEXT:    ds_write_b8 v1, v0 offset:2131; GFX7-NEXT:    v_mov_b32_e32 v0, s0132; GFX7-NEXT:    ds_write_b8 v1, v0 offset:3133; GFX7-NEXT:    s_bfe_u32 s3, s1, 0x80008134; GFX7-NEXT:    v_mov_b32_e32 v0, s1135; GFX7-NEXT:    s_lshr_b32 s0, s1, 16136; GFX7-NEXT:    ds_write_b8 v1, v0 offset:4137; GFX7-NEXT:    v_mov_b32_e32 v0, s3138; GFX7-NEXT:    ds_write_b8 v1, v0 offset:5139; GFX7-NEXT:    s_lshr_b32 s1, s1, 24140; GFX7-NEXT:    v_mov_b32_e32 v0, s0141; GFX7-NEXT:    ds_write_b8 v1, v0 offset:6142; GFX7-NEXT:    v_mov_b32_e32 v0, s1143; GFX7-NEXT:    ds_write_b8 v1, v0 offset:7144; GFX7-NEXT:    s_bfe_u32 s1, s2, 0x80008145; GFX7-NEXT:    v_mov_b32_e32 v0, s2146; GFX7-NEXT:    s_lshr_b32 s0, s2, 16147; GFX7-NEXT:    ds_write_b8 v1, v0 offset:8148; GFX7-NEXT:    v_mov_b32_e32 v0, s1149; GFX7-NEXT:    ds_write_b8 v1, v0 offset:9150; GFX7-NEXT:    s_lshr_b32 s1, s2, 24151; GFX7-NEXT:    v_mov_b32_e32 v0, s0152; GFX7-NEXT:    ds_write_b8 v1, v0 offset:10153; GFX7-NEXT:    v_mov_b32_e32 v0, s1154; GFX7-NEXT:    ds_write_b8 v1, v0 offset:11155; GFX7-NEXT:    s_endpgm156;157; GFX10-LABEL: store_lds_v3i32_align1:158; GFX10:       ; %bb.0:159; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10160; GFX10-NEXT:    s_waitcnt lgkmcnt(0)161; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x0162; GFX10-NEXT:    ; kill: killed $sgpr4_sgpr5163; GFX10-NEXT:    s_lshr_b32 s4, s0, 16164; GFX10-NEXT:    s_and_b32 s5, 0xffff, s0165; GFX10-NEXT:    v_mov_b32_e32 v0, s0166; GFX10-NEXT:    s_waitcnt lgkmcnt(0)167; GFX10-NEXT:    v_mov_b32_e32 v1, s3168; GFX10-NEXT:    s_lshr_b32 s0, s0, 24169; GFX10-NEXT:    s_lshr_b32 s3, s1, 16170; GFX10-NEXT:    s_and_b32 s6, 0xffff, s1171; GFX10-NEXT:    v_mov_b32_e32 v2, s1172; GFX10-NEXT:    s_lshr_b32 s1, s1, 24173; GFX10-NEXT:    s_and_b32 s8, 0xffff, s2174; GFX10-NEXT:    v_mov_b32_e32 v4, s4175; GFX10-NEXT:    v_mov_b32_e32 v5, s0176; GFX10-NEXT:    s_lshr_b32 s0, s6, 8177; GFX10-NEXT:    v_mov_b32_e32 v7, s1178; GFX10-NEXT:    s_lshr_b32 s1, s8, 8179; GFX10-NEXT:    s_lshr_b32 s7, s2, 16180; GFX10-NEXT:    v_mov_b32_e32 v3, s2181; GFX10-NEXT:    s_lshr_b32 s5, s5, 8182; GFX10-NEXT:    v_mov_b32_e32 v6, s3183; GFX10-NEXT:    v_mov_b32_e32 v9, s0184; GFX10-NEXT:    v_mov_b32_e32 v10, s1185; GFX10-NEXT:    s_lshr_b32 s0, s2, 24186; GFX10-NEXT:    v_mov_b32_e32 v8, s5187; GFX10-NEXT:    ds_write_b8 v1, v0188; GFX10-NEXT:    ds_write_b8 v1, v2 offset:4189; GFX10-NEXT:    ds_write_b8 v1, v4 offset:2190; GFX10-NEXT:    ds_write_b8 v1, v5 offset:3191; GFX10-NEXT:    ds_write_b8 v1, v6 offset:6192; GFX10-NEXT:    ds_write_b8 v1, v8 offset:1193; GFX10-NEXT:    ds_write_b8 v1, v9 offset:5194; GFX10-NEXT:    v_mov_b32_e32 v0, s7195; GFX10-NEXT:    v_mov_b32_e32 v2, s0196; GFX10-NEXT:    ds_write_b8 v1, v7 offset:7197; GFX10-NEXT:    ds_write_b8 v1, v3 offset:8198; GFX10-NEXT:    ds_write_b8 v1, v10 offset:9199; GFX10-NEXT:    ds_write_b8 v1, v0 offset:10200; GFX10-NEXT:    ds_write_b8 v1, v2 offset:11201; GFX10-NEXT:    s_endpgm202;203; GFX11-LABEL: store_lds_v3i32_align1:204; GFX11:       ; %bb.0:205; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10206; GFX11-NEXT:    s_waitcnt lgkmcnt(0)207; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0208; GFX11-NEXT:    s_and_b32 s5, 0xffff, s0209; GFX11-NEXT:    s_lshr_b32 s7, s2, 16210; GFX11-NEXT:    s_lshr_b32 s5, s5, 8211; GFX11-NEXT:    s_lshr_b32 s4, s0, 16212; GFX11-NEXT:    s_waitcnt lgkmcnt(0)213; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3214; GFX11-NEXT:    s_lshr_b32 s0, s0, 24215; GFX11-NEXT:    s_lshr_b32 s3, s1, 16216; GFX11-NEXT:    s_and_b32 s6, 0xffff, s1217; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s2218; GFX11-NEXT:    s_lshr_b32 s1, s1, 24219; GFX11-NEXT:    s_and_b32 s8, 0xffff, s2220; GFX11-NEXT:    v_dual_mov_b32 v8, s7 :: v_dual_mov_b32 v9, s5221; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s0222; GFX11-NEXT:    s_lshr_b32 s2, s2, 24223; GFX11-NEXT:    s_lshr_b32 s0, s6, 8224; GFX11-NEXT:    v_dual_mov_b32 v6, s3 :: v_dual_mov_b32 v7, s1225; GFX11-NEXT:    s_lshr_b32 s1, s8, 8226; GFX11-NEXT:    v_dual_mov_b32 v10, s2 :: v_dual_mov_b32 v11, s0227; GFX11-NEXT:    v_mov_b32_e32 v12, s1228; GFX11-NEXT:    ds_store_b8 v1, v0229; GFX11-NEXT:    ds_store_b8 v1, v9 offset:1230; GFX11-NEXT:    ds_store_b8 v1, v4 offset:2231; GFX11-NEXT:    ds_store_b8 v1, v5 offset:3232; GFX11-NEXT:    ds_store_b8 v1, v2 offset:4233; GFX11-NEXT:    ds_store_b8 v1, v11 offset:5234; GFX11-NEXT:    ds_store_b8 v1, v6 offset:6235; GFX11-NEXT:    ds_store_b8 v1, v7 offset:7236; GFX11-NEXT:    ds_store_b8 v1, v3 offset:8237; GFX11-NEXT:    ds_store_b8 v1, v12 offset:9238; GFX11-NEXT:    ds_store_b8 v1, v8 offset:10239; GFX11-NEXT:    ds_store_b8 v1, v10 offset:11240; GFX11-NEXT:    s_endpgm241  store <3 x i32> %x, ptr addrspace(3) %out, align 1242  ret void243}244 245define amdgpu_kernel void @store_lds_v3i32_align2(ptr addrspace(3) %out, <3 x i32> %x) {246; GFX9-LABEL: store_lds_v3i32_align2:247; GFX9:       ; %bb.0:248; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10249; GFX9-NEXT:    s_waitcnt lgkmcnt(0)250; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0251; GFX9-NEXT:    ; kill: killed $sgpr4_sgpr5252; GFX9-NEXT:    s_lshr_b32 s4, s0, 16253; GFX9-NEXT:    v_mov_b32_e32 v0, s0254; GFX9-NEXT:    s_waitcnt lgkmcnt(0)255; GFX9-NEXT:    v_mov_b32_e32 v1, s3256; GFX9-NEXT:    ds_write_b16 v1, v0257; GFX9-NEXT:    v_mov_b32_e32 v0, s4258; GFX9-NEXT:    ds_write_b16 v1, v0 offset:2259; GFX9-NEXT:    s_lshr_b32 s0, s1, 16260; GFX9-NEXT:    v_mov_b32_e32 v0, s1261; GFX9-NEXT:    ds_write_b16 v1, v0 offset:4262; GFX9-NEXT:    v_mov_b32_e32 v0, s0263; GFX9-NEXT:    ds_write_b16 v1, v0 offset:6264; GFX9-NEXT:    s_lshr_b32 s0, s2, 16265; GFX9-NEXT:    v_mov_b32_e32 v0, s2266; GFX9-NEXT:    ds_write_b16 v1, v0 offset:8267; GFX9-NEXT:    v_mov_b32_e32 v0, s0268; GFX9-NEXT:    ds_write_b16 v1, v0 offset:10269; GFX9-NEXT:    s_endpgm270;271; GFX7-LABEL: store_lds_v3i32_align2:272; GFX7:       ; %bb.0:273; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x4274; GFX7-NEXT:    s_waitcnt lgkmcnt(0)275; GFX7-NEXT:    s_load_dword s3, s[4:5], 0x0276; GFX7-NEXT:    s_mov_b32 m0, -1277; GFX7-NEXT:    s_lshr_b32 s4, s0, 16278; GFX7-NEXT:    v_mov_b32_e32 v0, s0279; GFX7-NEXT:    s_waitcnt lgkmcnt(0)280; GFX7-NEXT:    v_mov_b32_e32 v1, s3281; GFX7-NEXT:    ds_write_b16 v1, v0282; GFX7-NEXT:    v_mov_b32_e32 v0, s4283; GFX7-NEXT:    ds_write_b16 v1, v0 offset:2284; GFX7-NEXT:    s_lshr_b32 s0, s1, 16285; GFX7-NEXT:    v_mov_b32_e32 v0, s1286; GFX7-NEXT:    ds_write_b16 v1, v0 offset:4287; GFX7-NEXT:    v_mov_b32_e32 v0, s0288; GFX7-NEXT:    ds_write_b16 v1, v0 offset:6289; GFX7-NEXT:    s_lshr_b32 s0, s2, 16290; GFX7-NEXT:    v_mov_b32_e32 v0, s2291; GFX7-NEXT:    ds_write_b16 v1, v0 offset:8292; GFX7-NEXT:    v_mov_b32_e32 v0, s0293; GFX7-NEXT:    ds_write_b16 v1, v0 offset:10294; GFX7-NEXT:    s_endpgm295;296; GFX10-LABEL: store_lds_v3i32_align2:297; GFX10:       ; %bb.0:298; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10299; GFX10-NEXT:    s_waitcnt lgkmcnt(0)300; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x0301; GFX10-NEXT:    ; kill: killed $sgpr4_sgpr5302; GFX10-NEXT:    v_mov_b32_e32 v0, s0303; GFX10-NEXT:    s_waitcnt lgkmcnt(0)304; GFX10-NEXT:    v_mov_b32_e32 v1, s3305; GFX10-NEXT:    s_lshr_b32 s4, s0, 16306; GFX10-NEXT:    v_mov_b32_e32 v2, s1307; GFX10-NEXT:    s_lshr_b32 s0, s1, 16308; GFX10-NEXT:    v_mov_b32_e32 v3, s2309; GFX10-NEXT:    s_lshr_b32 s1, s2, 16310; GFX10-NEXT:    v_mov_b32_e32 v4, s4311; GFX10-NEXT:    v_mov_b32_e32 v5, s0312; GFX10-NEXT:    v_mov_b32_e32 v6, s1313; GFX10-NEXT:    ds_write_b16 v1, v0314; GFX10-NEXT:    ds_write_b16 v1, v2 offset:4315; GFX10-NEXT:    ds_write_b16 v1, v3 offset:8316; GFX10-NEXT:    ds_write_b16 v1, v4 offset:2317; GFX10-NEXT:    ds_write_b16 v1, v5 offset:6318; GFX10-NEXT:    ds_write_b16 v1, v6 offset:10319; GFX10-NEXT:    s_endpgm320;321; GFX11-LABEL: store_lds_v3i32_align2:322; GFX11:       ; %bb.0:323; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10324; GFX11-NEXT:    s_waitcnt lgkmcnt(0)325; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0326; GFX11-NEXT:    s_lshr_b32 s4, s0, 16327; GFX11-NEXT:    s_waitcnt lgkmcnt(0)328; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3329; GFX11-NEXT:    s_lshr_b32 s0, s1, 16330; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s4331; GFX11-NEXT:    s_lshr_b32 s1, s2, 16332; GFX11-NEXT:    v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v5, s0333; GFX11-NEXT:    v_mov_b32_e32 v6, s1334; GFX11-NEXT:    ds_store_b16 v1, v0335; GFX11-NEXT:    ds_store_b16 v1, v3 offset:2336; GFX11-NEXT:    ds_store_b16 v1, v2 offset:4337; GFX11-NEXT:    ds_store_b16 v1, v5 offset:6338; GFX11-NEXT:    ds_store_b16 v1, v4 offset:8339; GFX11-NEXT:    ds_store_b16 v1, v6 offset:10340; GFX11-NEXT:    s_endpgm341  store <3 x i32> %x, ptr addrspace(3) %out, align 2342  ret void343}344 345define amdgpu_kernel void @store_lds_v3i32_align4(ptr addrspace(3) %out, <3 x i32> %x) {346; GFX9-LABEL: store_lds_v3i32_align4:347; GFX9:       ; %bb.0:348; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10349; GFX9-NEXT:    s_waitcnt lgkmcnt(0)350; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0351; GFX9-NEXT:    ; kill: killed $sgpr4_sgpr5352; GFX9-NEXT:    v_mov_b32_e32 v0, s0353; GFX9-NEXT:    s_waitcnt lgkmcnt(0)354; GFX9-NEXT:    v_mov_b32_e32 v1, s3355; GFX9-NEXT:    v_mov_b32_e32 v2, s1356; GFX9-NEXT:    v_mov_b32_e32 v3, s2357; GFX9-NEXT:    ds_write2_b32 v1, v0, v2 offset1:1358; GFX9-NEXT:    ds_write_b32 v1, v3 offset:8359; GFX9-NEXT:    s_endpgm360;361; GFX7-LABEL: store_lds_v3i32_align4:362; GFX7:       ; %bb.0:363; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x4364; GFX7-NEXT:    s_waitcnt lgkmcnt(0)365; GFX7-NEXT:    s_load_dword s3, s[4:5], 0x0366; GFX7-NEXT:    s_mov_b32 m0, -1367; GFX7-NEXT:    v_mov_b32_e32 v0, s0368; GFX7-NEXT:    s_waitcnt lgkmcnt(0)369; GFX7-NEXT:    v_mov_b32_e32 v1, s3370; GFX7-NEXT:    v_mov_b32_e32 v2, s1371; GFX7-NEXT:    ds_write2_b32 v1, v0, v2 offset1:1372; GFX7-NEXT:    v_mov_b32_e32 v0, s2373; GFX7-NEXT:    ds_write_b32 v1, v0 offset:8374; GFX7-NEXT:    s_endpgm375;376; GFX10-LABEL: store_lds_v3i32_align4:377; GFX10:       ; %bb.0:378; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10379; GFX10-NEXT:    s_waitcnt lgkmcnt(0)380; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x0381; GFX10-NEXT:    ; kill: killed $sgpr4_sgpr5382; GFX10-NEXT:    v_mov_b32_e32 v0, s0383; GFX10-NEXT:    s_waitcnt lgkmcnt(0)384; GFX10-NEXT:    v_mov_b32_e32 v1, s3385; GFX10-NEXT:    v_mov_b32_e32 v2, s1386; GFX10-NEXT:    v_mov_b32_e32 v3, s2387; GFX10-NEXT:    ds_write2_b32 v1, v0, v2 offset1:1388; GFX10-NEXT:    ds_write_b32 v1, v3 offset:8389; GFX10-NEXT:    s_endpgm390;391; GFX11-LABEL: store_lds_v3i32_align4:392; GFX11:       ; %bb.0:393; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10394; GFX11-NEXT:    s_waitcnt lgkmcnt(0)395; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0396; GFX11-NEXT:    s_waitcnt lgkmcnt(0)397; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3398; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s2399; GFX11-NEXT:    ds_store_2addr_b32 v1, v0, v2 offset1:1400; GFX11-NEXT:    ds_store_b32 v1, v3 offset:8401; GFX11-NEXT:    s_endpgm402  store <3 x i32> %x, ptr addrspace(3) %out, align 4403  ret void404}405 406define amdgpu_kernel void @store_lds_v3i32_align8(ptr addrspace(3) %out, <3 x i32> %x) {407; GFX9-LABEL: store_lds_v3i32_align8:408; GFX9:       ; %bb.0:409; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10410; GFX9-NEXT:    s_waitcnt lgkmcnt(0)411; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0412; GFX9-NEXT:    ; kill: killed $sgpr4_sgpr5413; GFX9-NEXT:    v_mov_b32_e32 v0, s0414; GFX9-NEXT:    s_waitcnt lgkmcnt(0)415; GFX9-NEXT:    v_mov_b32_e32 v1, s3416; GFX9-NEXT:    v_mov_b32_e32 v2, s1417; GFX9-NEXT:    v_mov_b32_e32 v3, s2418; GFX9-NEXT:    ds_write2_b32 v1, v0, v2 offset1:1419; GFX9-NEXT:    ds_write_b32 v1, v3 offset:8420; GFX9-NEXT:    s_endpgm421;422; GFX7-LABEL: store_lds_v3i32_align8:423; GFX7:       ; %bb.0:424; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x4425; GFX7-NEXT:    s_waitcnt lgkmcnt(0)426; GFX7-NEXT:    s_load_dword s3, s[4:5], 0x0427; GFX7-NEXT:    s_mov_b32 m0, -1428; GFX7-NEXT:    v_mov_b32_e32 v0, s0429; GFX7-NEXT:    s_waitcnt lgkmcnt(0)430; GFX7-NEXT:    v_mov_b32_e32 v1, s3431; GFX7-NEXT:    v_mov_b32_e32 v2, s1432; GFX7-NEXT:    ds_write2_b32 v1, v0, v2 offset1:1433; GFX7-NEXT:    v_mov_b32_e32 v0, s2434; GFX7-NEXT:    ds_write_b32 v1, v0 offset:8435; GFX7-NEXT:    s_endpgm436;437; GFX10-LABEL: store_lds_v3i32_align8:438; GFX10:       ; %bb.0:439; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10440; GFX10-NEXT:    s_waitcnt lgkmcnt(0)441; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x0442; GFX10-NEXT:    ; kill: killed $sgpr4_sgpr5443; GFX10-NEXT:    v_mov_b32_e32 v0, s0444; GFX10-NEXT:    s_waitcnt lgkmcnt(0)445; GFX10-NEXT:    v_mov_b32_e32 v1, s3446; GFX10-NEXT:    v_mov_b32_e32 v2, s1447; GFX10-NEXT:    v_mov_b32_e32 v3, s2448; GFX10-NEXT:    ds_write2_b32 v1, v0, v2 offset1:1449; GFX10-NEXT:    ds_write_b32 v1, v3 offset:8450; GFX10-NEXT:    s_endpgm451;452; GFX11-LABEL: store_lds_v3i32_align8:453; GFX11:       ; %bb.0:454; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10455; GFX11-NEXT:    s_waitcnt lgkmcnt(0)456; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0457; GFX11-NEXT:    s_waitcnt lgkmcnt(0)458; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3459; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s2460; GFX11-NEXT:    ds_store_2addr_b32 v1, v0, v2 offset1:1461; GFX11-NEXT:    ds_store_b32 v1, v3 offset:8462; GFX11-NEXT:    s_endpgm463  store <3 x i32> %x, ptr addrspace(3) %out, align 8464  ret void465}466 467define amdgpu_kernel void @store_lds_v3i32_align16(ptr addrspace(3) %out, <3 x i32> %x) {468; GFX9-LABEL: store_lds_v3i32_align16:469; GFX9:       ; %bb.0:470; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10471; GFX9-NEXT:    s_waitcnt lgkmcnt(0)472; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0473; GFX9-NEXT:    ; kill: killed $sgpr4_sgpr5474; GFX9-NEXT:    v_mov_b32_e32 v0, s0475; GFX9-NEXT:    v_mov_b32_e32 v1, s1476; GFX9-NEXT:    v_mov_b32_e32 v2, s2477; GFX9-NEXT:    s_waitcnt lgkmcnt(0)478; GFX9-NEXT:    v_mov_b32_e32 v3, s3479; GFX9-NEXT:    ds_write_b96 v3, v[0:2]480; GFX9-NEXT:    s_endpgm481;482; GFX7-LABEL: store_lds_v3i32_align16:483; GFX7:       ; %bb.0:484; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x4485; GFX7-NEXT:    s_waitcnt lgkmcnt(0)486; GFX7-NEXT:    s_load_dword s3, s[4:5], 0x0487; GFX7-NEXT:    s_mov_b32 m0, -1488; GFX7-NEXT:    v_mov_b32_e32 v0, s0489; GFX7-NEXT:    v_mov_b32_e32 v1, s1490; GFX7-NEXT:    v_mov_b32_e32 v2, s2491; GFX7-NEXT:    s_waitcnt lgkmcnt(0)492; GFX7-NEXT:    v_mov_b32_e32 v3, s3493; GFX7-NEXT:    ds_write_b96 v3, v[0:2]494; GFX7-NEXT:    s_endpgm495;496; GFX10-LABEL: store_lds_v3i32_align16:497; GFX10:       ; %bb.0:498; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10499; GFX10-NEXT:    s_waitcnt lgkmcnt(0)500; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x0501; GFX10-NEXT:    ; kill: killed $sgpr4_sgpr5502; GFX10-NEXT:    v_mov_b32_e32 v0, s0503; GFX10-NEXT:    v_mov_b32_e32 v1, s1504; GFX10-NEXT:    v_mov_b32_e32 v2, s2505; GFX10-NEXT:    s_waitcnt lgkmcnt(0)506; GFX10-NEXT:    v_mov_b32_e32 v3, s3507; GFX10-NEXT:    ds_write_b96 v3, v[0:2]508; GFX10-NEXT:    s_endpgm509;510; GFX11-LABEL: store_lds_v3i32_align16:511; GFX11:       ; %bb.0:512; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10513; GFX11-NEXT:    s_waitcnt lgkmcnt(0)514; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0515; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1516; GFX11-NEXT:    s_waitcnt lgkmcnt(0)517; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3518; GFX11-NEXT:    ds_store_b96 v3, v[0:2]519; GFX11-NEXT:    s_endpgm520  store <3 x i32> %x, ptr addrspace(3) %out, align 16521  ret void522}523